NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

NeoHorse-1. Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness. By NeoHorse Team and thirty-six others. Published on the eighth of September. Its code repository is public and real: about thirty-five source files.

Recursive self-improvement, or RSI, needs a concrete mechanism: a way for an AI system to observe its own capabilities and turn that evidence into its next round of learning. Without one, self-improvement stays an aspiration rather than a process. The mechanism has to specify what is observed, how it becomes training data, and how the next round is chosen.

NeoHorse-1 is a family of agent-native models built to explore that path through agentic post-training. The system combines a heterogeneous pool of models with intelligent routing. For each user turn, it records the capability demand it predicted, the service tier it selected, and the interaction that followed. Those records are the evidence the system learns from. Routing is not only a way of serving requests; each decision it makes is also a record of what capability a turn seemed to need.

The records are converted into training examples that preserve interleaved reasoning, tool calls, and the context of the harness. Admission is gated three ways: structural validation, semantic evaluation along six dimensions, and labelling at the level of subscenes. Routing signals then organise supervised fine-tuning into a three-stage curriculum. They extend to routing- guided on-policy distillation, in which a teacher supervises the student's own responses under the same progression. The curriculum and the distillation follow the same ordering, both derived from what routing observed.

Capability-guided allocation closes the loop. It converts evaluation feedback into the next training mixture. That makes a cycle of evaluation, selection, and update, in which what the system learns to do shapes what it learns from next. Each round's training mixture is chosen by the previous round's results.

Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from fifty-eight point nine four to sixty-four point eight seven for the four billion parameter model. For the nine billion model it rises from sixty-five point six to sixty- nine point zero four. That substantially narrows the aggregate gap between the post-trained four billion model and the nine billion base model. Both sizes improve, with the larger gain at the smaller size.

The authors frame this as an initial prototype of a feedback-driven process, with a path toward RSI mediated by the harness over successive iterations. What it demonstrates is the loop itself: routing produces records, records become training data, evaluation steers the next mixture, and the smaller model moves toward the larger one.